Cable cover plate anomaly detection method and system based on variational auto-encoder

Abstract

The invention relates to the technical field of cable cover plate detection, and discloses a cable cover plate anomaly detection method and system based on a variational auto-encoder, and the method comprises the steps: carrying out the denoising processing of the original data of a cable cover plate collected by a sensor, obtaining a denoised signal, carrying out the dynamic resampling of the denoised signal according to a local change rate, and carrying out the detection of the anomaly of the cable cover plate. Resampling data is obtained; inputting the resampling data into a pre-established data reconstruction model to obtain reconstructed data, the data reconstruction model being constructed based on a variational auto-encoder and being obtained by training based on a dynamic learning rate adjustment strategy and a data enhancement strategy; and performing anomaly detection according to the reconstruction error change rate of the reconstruction data to obtain abnormal data, and generating a corresponding alarm prompt according to the abnormal data. According to the method, the integrity and reliability of data can be improved, the flexibility and accuracy of anomaly detection are improved, the change of the operation state of the cable cover plate can be responded in time, and the real-time performance and effectiveness of anomaly recognition are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of cable cover plate detection, and in particular to a cable cover plate abnormality detection method and system based on variational autoencoder. Background Art

[0002] Current cable cover monitoring and anomaly detection technologies mostly rely on traditional data processing methods, mainly including basic techniques such as signal filtering and statistical analysis. However, these methods have a series of problems and limitations in complex practical application environments.

[0003] First, the sensor data in the cable cover is often affected by the external environment during the collection process, such as temperature fluctuations, mechanical vibrations, and electromagnetic interference, resulting in a lot of noise in the collected data. Traditional signal filtering methods usually use fixed parameters, which can remove some noise but may also lead to the loss of key information. In the face of sudden failures or progressive damage, the filter may not be able to identify signal changes in a timely and accurate manner, thus affecting the accuracy of anomaly detection.

[0004] Secondly, most of the existing anomaly detection methods are based on simple rule settings and fixed threshold judgments, which have relatively good detection effects under normal operating conditions, but show poor adaptability in complex dynamic environments. This type of method is prone to two extreme situations: one is the failure to detect abnormal information in a timely manner, and the other is that the false alarm rate is too high, which affects the stability of the system. In addition, with the increase in the number of sensors in the cable cover system, the amount of data becomes increasingly large and complex. Traditional analysis methods have low computational efficiency when processing massive data and cannot meet real-time requirements.

[0005] Furthermore, traditional anomaly detection technology often ignores the correlation and multi-dimensional characteristics between data, resulting in insufficient ability to identify abnormal patterns. In some cases, the system cannot effectively consider multiple sensor data to form a comprehensive fault analysis. This limitation not only limits the accuracy of anomaly detection, but also reduces the ability to predict faults.

[0006] Finally, existing technologies mostly use linear models or simple machine learning algorithms, which are difficult to adapt to the increasingly complex cable cover monitoring needs. These methods lack the feature extraction capabilities of deep learning and cannot automatically obtain important features from raw data, resulting in insufficient recognition of complex patterns.

[0007] In summary, traditional cable cover monitoring and anomaly detection technologies have obvious shortcomings in noise processing, data analysis, real-time and complex pattern recognition, and it is urgent to develop more advanced technical means to improve the accuracy and efficiency of monitoring. Summary of the invention

[0008] In order to solve the above technical problems, the present invention provides a cable cover anomaly detection method and system based on variational autoencoder, so as to solve the problems existing in traditional cable cover monitoring and anomaly detection technologies in noise processing, data analysis, real-time and complex pattern recognition, so as to achieve the effect of improving the real-time and effectiveness of anomaly identification.

[0009] In a first aspect, the present invention provides a cable cover abnormality detection method based on a variational autoencoder, the method comprising: De-noising the original data of the cable cover plate collected by the sensor to obtain a de-noised signal, and dynamically re-sampling the de-noised signal according to the local change rate to obtain re-sampled data; Inputting the resampled data into a pre-established data reconstruction model to obtain reconstructed data, wherein the data reconstruction model is constructed based on a variational autoencoder and trained based on a dynamic learning rate adjustment strategy and a data enhancement strategy; Performing anomaly detection according to a reconstruction error change rate of the reconstructed data to obtain abnormal data, and generating a corresponding alarm reminder according to the abnormal data; Among them, the data reconstruction model consists of an encoder and a corresponding decoder. The encoder is constructed based on a deep convolutional neural network, and a dynamic adjustment mechanism is used to adjust the mean and variance of the generated latent variables. The decoder is composed of a deconvolution layer, and an adaptive feature fusion mechanism and a multidimensional recovery mechanism are used to reconstruct the latent variables.

[0010] Furthermore, the step of performing denoising processing on the raw data of the cable cover plate collected by the sensor to obtain a denoised signal includes: Collecting raw data of the cable cover through sensors, wherein the raw data includes temperature data, humidity data and pressure data; Performing time-frequency conversion on the original data by Fourier transform to obtain a frequency domain signal, calculating the energy spectrum density of the frequency domain signal, comparing the energy spectrum density with a density threshold, and filtering interference signals on the frequency domain signal according to the comparison result to obtain a first frequency domain signal, wherein the density threshold is determined by the median and standard deviation of the energy spectrum density; The first frequency domain signal is frequency filtered using a frequency filter constructed based on a window function to obtain a second frequency domain signal, and the second frequency domain signal is time-frequency inverted to obtain a denoised signal.

[0011] Furthermore, the step of dynamically resampling the denoised signal according to the local change rate to obtain resampled data includes: Dividing the denoised signal into a plurality of time series intervals, and calculating the local change rate of the denoised signal according to the signal difference between adjacent sampling points in each time series interval; Comparing the local change rate with a change rate threshold, and determining a data sampling mode for each sensor according to the comparison result, wherein the data sampling mode includes a high-frequency sampling mode and a low-frequency sampling mode; Determine a resampling weight of each sensor based on historical data of each sensor, and dynamically adjust the sampling frequency in the data sampling mode according to the resampling weight to obtain a dynamic resampling frequency; According to the dynamic resampling frequency, the denoised signal corresponding to each sensor is dynamically resampled to obtain resampled data.

[0012] Furthermore, the encoder includes a plurality of convolutional layers, a plurality of pooling layers and a first fully connected layer, the number of convolutional layers and pooling layers is the same, the convolutional kernel sizes used in each convolutional layer are different, the output data of each convolutional layer is the input data of the corresponding pooling layer, the output data of the corresponding pooling layer is the input data of the next convolutional layer, and the output data of the last pooling layer is the input data of the first fully connected layer; The encoder uses a dynamic adjustment mechanism based on the rate of change to adjust the mean and variance of the generated latent variables; The decoder includes a plurality of sequentially connected deconvolution layers and a second fully connected layer, the deconvolution kernel sizes used in the various deconvolution layers are different, the last deconvolution layer is connected to the second fully connected layer, the number of deconvolution layers is consistent with the number of convolution layers of the encoder, and each deconvolution layer and the corresponding convolution layer use a shared weight matrix; The encoder and the decoder have a skip connection structure, and the decoder reconstructs the latent variables using a multi-dimensional recovery mechanism combining the mean and the variance.

[0013] Furthermore, the loss function of the data reconstruction model is a weighted function based on reconstruction error loss and latent space regularization loss, and the weight coefficient in the loss function is a dynamic weight coefficient that is dynamically adjusted based on the reconstruction error change rate and the latent space distribution change rate; The loss function is expressed by the following formula: In the formula, represents the reconstruction error loss, represents the latent space regularization loss, represents the weight coefficient; The weight coefficient is expressed by the following formula: In the formula, represents the adjusted weight coefficient, represents the weight coefficient before adjustment, represents the rate of change of reconstruction error, Represents the rate of change of the potential spatial distribution.

[0014] Furthermore, the steps of training the data reconstruction model based on a dynamic learning rate adjustment strategy include: Calculating the rate of change and the gradient norm of the loss function, updating the learning rate according to the rate of change and the gradient norm, and iteratively training the data reconstruction model using the updated learning rate; The learning rate is expressed by the following formula: In the formula, represents the new learning rate, represents the learning rate of the previous iteration, represents the loss change rate, represents the gradient norm, Indicates control parameters.

[0015] Furthermore, the steps of training the data reconstruction model based on a dynamic learning rate adjustment strategy include: The learning rate is updated according to the current number of training steps, the total number of steps in the learning rate cycle and the extreme value of the learning rate, and the updated learning rate is used to iteratively train the data reconstruction model; The learning rate is expressed by the following formula: in, represents the new learning rate, and Respectively represent the minimum and maximum values ​​of the learning rate, d represents the current number of training steps, Represents the total number of steps in the learning rate cycle.

[0016] Furthermore, the steps of training the data reconstruction model based on the data enhancement strategy include: Acquire historical data of the cable cover plate collected by the sensor, and perform frequency domain disturbance enhancement and time domain random masking enhancement on the historical data to obtain first enhanced historical data and second enhanced historical data; A data set is constructed according to the historical data, the first enhanced historical data and the second enhanced historical data, and the data set is used to train the data reconstruction model.

[0017] Further, the step of performing anomaly detection according to the reconstruction error change rate of the reconstructed data to obtain abnormal data, and generating a corresponding alarm reminder according to the abnormal data includes: Obtaining a reconstruction error change rate according to a reconstruction error between the original data and the reconstructed data, comparing the reconstruction error change rate with an abnormal threshold, and obtaining abnormal data according to a comparison result; According to the duration and degree of abnormality of the abnormal data, an alarm reminder of a corresponding level is generated.

[0018] In a second aspect, the present invention provides a cable cover plate abnormality detection system based on a variational autoencoder, the system comprising: The data sampling module is used to perform denoising on the original data of the cable cover plate collected by the sensor to obtain a denoised signal, and dynamically resample the denoised signal according to the local change rate to obtain resampled data; A data reconstruction module, used for inputting the resampled data into a pre-established data reconstruction model to obtain reconstructed data, wherein the data reconstruction model is constructed based on a variational autoencoder and trained based on a dynamic learning rate adjustment strategy and a data enhancement strategy; The data reconstruction model consists of an encoder and a corresponding decoder. The encoder is constructed based on a deep convolutional neural network and uses a dynamic adjustment mechanism to adjust the mean and variance of the generated latent variables. The decoder consists of a deconvolution layer and uses an adaptive feature fusion mechanism and a multi-dimensional recovery mechanism to reconstruct the latent variables. The anomaly detection module is used to perform anomaly detection according to the reconstruction error change rate of the reconstructed data to obtain abnormal data, and generate corresponding alarm reminders according to the abnormal data.

[0019] The present invention provides a cable cover plate anomaly detection method and system based on variational autoencoder. The present invention can realize data reconstruction and fusion and dynamic identification of abnormal states through the generative modeling ability, unsupervised learning ability and automatic learning ability of the variational autoencoder, which can not only improve the integrity and reliability of data, improve the flexibility and accuracy of anomaly detection, but also can respond to the changes in the operating state of the cable cover plate in a timely manner, thereby improving the real-time and effectiveness of anomaly identification. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 is a flow chart of a cable cover plate abnormality detection method based on a variational autoencoder in an embodiment of the present invention; Figure 2 It is a structural schematic diagram of a cable cover anomaly detection system based on a variational autoencoder in an embodiment of the present invention. DETAILED DESCRIPTION

[0021] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0022] See also Figure 1 The first embodiment of the present invention proposes a cable cover abnormality detection method based on a variational autoencoder, which includes steps S10 to S30: Step S10, performing denoising processing on the original data of the cable cover plate collected by the sensor to obtain a denoised signal, and dynamically resampling the denoised signal according to the local change rate to obtain resampled data; Step S20, inputting the resampled data into a pre-established data reconstruction model to obtain reconstructed data, wherein the data reconstruction model is constructed based on a variational autoencoder and trained based on a dynamic learning rate adjustment strategy and a data enhancement strategy; Step S30, performing anomaly detection according to the reconstruction error change rate of the reconstructed data to obtain abnormal data, and generating a corresponding alarm reminder according to the abnormal data.

[0023] The present invention determines whether the cable cover is abnormal by analyzing the data collected by the sensor installed on the cable cover. Since the data collected by the sensor on the cable cover is usually affected by environmental noise, it is very necessary to filter the noise before data analysis. The specific steps of noise filtering are as follows: Collecting raw data of the cable cover through sensors, wherein the raw data includes temperature data, humidity data and pressure data; Performing time-frequency conversion on the original data by Fourier transform to obtain a frequency domain signal, calculating the energy spectrum density of the frequency domain signal, comparing the energy spectrum density with a density threshold, and filtering interference signals on the frequency domain signal according to the comparison result to obtain a first frequency domain signal, wherein the density threshold is determined by the median and standard deviation of the energy spectrum density; The first frequency domain signal is frequency filtered using a frequency filter constructed based on a window function to obtain a second frequency domain signal, and the second frequency domain signal is time-frequency inverted to obtain a denoised signal.

[0024] In this embodiment, the relevant data of the cable cover is collected according to multiple sensors on the cable cover, including temperature data, humidity data, pressure data and other data. The original data contains the real-time information generated by the cable cover in actual operation, which is the basis of the entire data processing. Therefore, the collection of these data is the premise of subsequent signal processing. For these original data, the spectrum separation technology is used to convert the data from the time domain to the frequency domain through Fourier transform. Assume that the original data collected from N sensors on the cable cover is , represents the time domain signal of the i-th sensor, where t is the time variable and i is the sensor number. The time domain signal is converted into a frequency domain signal using Fourier transform : in, is the frequency domain signal, is the frequency, Is an imaginary unit.

[0025] The reason for converting data from the time domain to the frequency domain is that the periodic characteristics and frequency distribution of the signal can be presented more clearly in the frequency domain, making the subsequent filtering process more accurate. Therefore, the frequency domain analysis can more effectively identify the different frequency components in the signal and help distinguish noise from useful signals. In this embodiment, a noise removal technology based on energy spectral density is adopted. Specifically, the energy spectral density PSD of the frequency domain signal is calculated by the following formula: in, is the total time length of the signal.

[0026] The energy spectral density PSD can effectively reflect the energy distribution of the signal at each frequency. Therefore, by setting a density threshold and comparing the PSD with the density threshold, we can identify the frequency components with higher energy that do not meet the signal characteristics by judging whether the energy of each frequency exceeds the threshold.

[0027] In a preferred embodiment, the present invention dynamically sets the density threshold by the median and standard deviation of the energy spectrum density. Specifically, the frequency domain signal is locally analyzed based on the sliding window. Assuming that the window length has N frequency points, such as N is 512, the median of the PSD is calculated in each sliding window to reflect the energy distribution, and the standard deviation of the PSD in the sliding window is calculated at the same time, and then the dynamic density threshold A is set according to the median and the standard deviation: Where k is the coefficient of control sensitivity, between 1.5 and 3, and M represents the median. 1 represents the standard deviation of the PSD.

[0028] The current PSD is compared with the density threshold A. If the PSD of a certain frequency point exceeds the density threshold, it is determined to be an abnormal signal. This embodiment traverses the entire frequency domain signal through a sliding window method, dynamically updates the median and standard deviation, and implements an adjustment threshold to adapt to changes in signal characteristics.

[0029] In this embodiment, the sliding window is used to analyze the frequency domain signal in order to localize the signal characteristics, so that a more detailed energy analysis can be performed in different frequency bands, and the changing characteristics of the signal in a short time can be captured, so that the detection process is more sensitive, thereby enhancing the accuracy of capturing the signal energy distribution and abnormal detection. In this embodiment, by calculating the median of the power spectral density (PSD) in each sliding window, the energy distribution of the frequency band can be more accurately reflected, thereby providing a benchmark for subsequent abnormal detection. At the same time, the standard deviation of the PSD is calculated in the same sliding window to evaluate the degree of change of the signal energy in the window. As a statistical indicator, the standard deviation can reveal the degree of discreteness of the energy distribution, and by setting the coefficient of controlling the sensitivity, the sensitivity and robustness are balanced, so that the identification of abnormal signals is more accurate, and false alarms or missed alarms caused by environmental changes can be effectively avoided. Finally, by traversing the sliding window, the median and standard deviation of the entire frequency domain signal are dynamically updated, and the threshold is adjusted in real time to continuously adapt to changes in signal characteristics. This embodiment not only improves the real-time and accuracy of anomaly detection, but also ensures the reliability of the detection method provided by the present invention under different operating conditions, thereby forming a closed-loop feedback mechanism in the entire data analysis process, making the detection method provided by the present invention more intelligent and flexible.

[0030] Through the above steps, potential interference signals can be effectively identified. After filtering the interference of the frequency domain signal, a first frequency domain signal is obtained. Then, the first frequency signal needs to be frequency filtered. In this embodiment, a Hamming window is selected as a window function to design a frequency filter. The Hamming window has a smooth spectrum characteristic, which can reduce sidelobe leakage and reduce the boundary effect of the frequency domain signal during the Fourier transform process, thereby improving the filtering effect. Of course, other window functions can also be used, which are only used as a preferred method and not specifically limited here.

[0031] The filtering range of the frequency filter, i.e., the lowest cutoff frequency and the highest cutoff frequency, can be determined according to the energy spectrum density. For example, the density range can be selected according to the energy spectrum density curve to determine the frequency range. The advantage of this is that only important frequency components in the signal can be retained. Then the frequency response value between the lowest cutoff frequency and the highest cutoff frequency is 1, which means that these frequency components will be retained, and the response value of other frequencies is 0, which means that these unnecessary frequency components will be filtered out. In addition, the cutoff frequency of the filter can be dynamically adjusted according to the change of the spectrum to ensure that the filter adapts to the changes in signal characteristics, thereby improving the denoising effect.

[0032] For the frequency domain signal after noise filtering, the frequency domain signal is converted into a time domain signal by inverse Fourier transform, so as to obtain the time domain signal after noise filtering, that is, the denoised signal. In the actual application environment, due to the complexity and uncertainty of the sensor signal change, in a preferred embodiment, the present invention improves the validity and processing efficiency of the data by dynamically resampling the denoised signal, and the specific steps of dynamic resampling include: Dividing the denoised signal into a plurality of time series intervals, and calculating the local change rate of the denoised signal according to the signal difference between adjacent sampling points in each time series interval; Comparing the local change rate with a change rate threshold, and determining a data sampling mode for each sensor according to the comparison result, wherein the data sampling mode includes a high-frequency sampling mode and a low-frequency sampling mode; Determine a resampling weight of each sensor based on historical data of each sensor, and dynamically adjust the sampling frequency in the data sampling mode according to the resampling weight to obtain a dynamic resampling frequency; According to the dynamic resampling frequency, the denoised signal corresponding to each sensor is dynamically resampled to obtain resampled data.

[0033] In this embodiment, the denoised signal is first divided into continuous time intervals, that is, multiple time series intervals, and then the local change rate of the denoised signal is calculated in each interval. The local change rate here is calculated by the signal difference between adjacent sampling points: in, represents the local rate of change of the i-th sensor at time t, represents the sensor signal after denoising, The sampling period is the time period during which the drastic fluctuation or slow change of the signal can be captured by calculating the local rate of change, thereby being able to identify possible abnormal changes in a timely manner and provide a basis for subsequent resampling.

[0034] According to the size of the local change rate, this embodiment adopts an adaptive resampling strategy. When the local change rate exceeds the set threshold, it switches to a high-frequency sampling mode to obtain more sampled data when the signal fluctuates violently, ensuring that key features can be captured; conversely, when the local change rate is lower than the set threshold, it switches to a low-frequency sampling mode to reduce unnecessary data collection and processing burdens. By dynamically adjusting the sampling frequency, the adaptability of the method provided by the present invention in different situations can be improved.

[0035] Furthermore, since different sensors have different importance, in this embodiment, on the basis of dynamic resampling, a weight mechanism is added according to the importance of sensor data to ensure that key features are not ignored during the resampling process and to adapt to data changes of the cable cover under different operating conditions.

[0036] In this embodiment, the importance of the data collected by the sensor is measured based on the mean and standard deviation of the historical data, and the calculation method of the resampling weight is as follows: in, is the historical signal mean of sensor i, is the standard deviation of the signal.

[0037] The resampling weight can reflect the contribution of each sensor in the anomaly detection process. Through the weight mechanism, sensors with drastic changes and large contributions to abnormal signals can be processed with higher priority to ensure that key data is not ignored by low-frequency sampling.

[0038] Finally, the resampling frequency is dynamically adjusted by calculating the resampling weight in real time. The dynamic resampling frequency is calculated using the following formula: in, is the basic sampling frequency, is the weighting coefficient.

[0039] For the above-mentioned high-frequency sampling mode and low-frequency sampling mode, there are different basic sampling frequencies respectively. In this embodiment, the basic sampling frequency is adjusted by resampling weights, which provides a flexible response strategy for signal changes. When the signal changes drastically, a higher resampling frequency can ensure the capture of key features; when the signal is stable, reducing the sampling frequency can reduce unnecessary data processing burden. This dynamic adjustment mechanism not only improves the efficiency of data processing, but also effectively improves the accuracy and reliability of the method provided by the present invention in the process of anomaly detection.

[0040] For the resampled data, the present invention reconstructs the resampled data through a data reconstruction model. In a preferred embodiment, the data reconstruction model is constructed using a variational autoencoder. The structure of the data reconstruction model in this embodiment and its training steps are described in detail below.

[0041] The data reconstruction model in this embodiment is composed of an encoder and a corresponding decoder. The encoder is constructed using a deep convolutional neural network, including multiple convolutional layers and pooling layers. In the convolutional layer design of the encoder, multi-scale convolutional kernels are used for feature extraction. Each layer uses convolutional kernels of different sizes to capture multi-scale feature changes in the cable cover data, thereby extracting local and global feature information at the same time to improve the accuracy of anomaly detection. The convolutional layer outputs The calculation formula is: in, represents the convolution output of layer l, represents the convolution kernel weight matrix of the lth layer, is the convolution operation, is the input data or the output of the previous layer, is the bias term, is the activation function ReLU.

[0042] An adaptive pooling layer is introduced after each convolution layer. The adaptive pooling layer dynamically adjusts the pooling window size according to the input signal changes to ensure that different downsampling rates are used in different areas, thereby reducing data redundancy while retaining important features. The output of the pooling layer is The calculation formula is: in, is the adaptive pooling operation, and They are the height and width of the adaptively adjusted pooling window.

[0043] The final convolutional pooling output of the encoder is mapped to the latent space through a fully connected layer to generate the mean of the latent variable and variance , thereby capturing the potential characteristic structure of the data, the formula for generating the mean and variance is as follows: in, is the mean in the latent space, is the variance in the latent space, is used to generate the mean in the latent space The weight matrix of is used to generate the variance in the latent space The weight matrix of and is the corresponding bias term, It is the feature vector output by the final convolutional pooling layer.

[0044] In this embodiment, a multi-scale convolution kernel is used for feature extraction. By using convolution kernels of different sizes in each layer, the encoder can capture the multi-scale feature changes in the cable cover data. This strategy not only enables local features and global features to be extracted at the same time, thereby improving the accuracy of anomaly detection, but also increases the sensitivity of the model to different feature scales in the data. During the calculation process, the formula of the convolution output clearly expresses the relationship between the convolution kernel weight matrix of each layer and the input data, and combines the bias term and the activation function ReLU to ensure the expression ability of nonlinear features. In addition, an adaptive pooling layer is introduced after each convolution layer to dynamically adjust the pooling window size according to the changes in the input signal. This method ensures that different downsampling rates are used in different areas, which can effectively retain important features and reduce data redundancy. Through such dynamic adjustment, the encoder can better adapt to the different features of the input signal, thereby improving the effectiveness of subsequent processing. The formula of the pooling layer output further clarifies the mechanism of the adaptive pooling operation and reflects the dynamic adjustment ability of the height and width of the pooling window. The final convolutional pooling output of the encoder is mapped to the latent space through a fully connected layer to generate the mean and variance of the latent variables, which can effectively capture the potential feature structure of the data and provide an important basis for subsequent anomaly detection.

[0045] In order to make the model adapt to sudden abnormal signals in the data, so as to timely adjust the mapping of the latent space when the cable cover sensor data changes dramatically, in a preferred embodiment, the present invention also provides a dynamic adjustment mechanism for adjusting the mean and variance in the latent space based on data changes. In this mechanism, the following formula is used to adjust the mean and variance: in, and are the adjusted mean and variance, and are the adjustment coefficients, and are the rates of change of the mean and variance in the latent space, respectively.

[0046] In the adjustment formula of this embodiment, the rate of change is introduced into the adjusted mean and variance, ensuring the flexibility and adaptability of the model. This dynamic adjustment mechanism not only enhances the model's ability to respond to emergencies, but also improves the robustness of the overall system.

[0047] In addition to the above-mentioned encoder, the data reconstruction model in this embodiment also includes a decoder corresponding to the encoder, wherein the decoder includes multiple deconvolution layers, and the number of deconvolution layers is consistent with the number of convolution layers in the encoder, and the two are in a one-to-one correspondence, thereby ensuring that the reconstruction of the latent variables is consistent with the dimension of the input data, and each deconvolution layer and the corresponding convolution layer use a shared weight matrix to ensure the continuity and consistency of the feature mapping, thereby ensuring that the features extracted by the encoder can be used in the reconstruction process, which helps to retain key feature information during reconstruction.

[0048] Each layer of deconvolution operation is accompanied by a specific activation function Leaky ReLU, which is used to gradually restore the high-order features in the latent space at each layer. The deconvolution operation gradually increases the size of the feature map through the stride operation, and restores it from the latent space to the same spatial size as the input data. The deconvolution process is described by the formula as follows: in, is the weight matrix of the encoder convolutional layer The transpose of is used for deconvolution operation, is the input of the decoder layer l, that is, the output of the previous layer, is the bias term of the deconvolution layer, is the activation function Leaky ReLU.

[0049] In this embodiment, in the reconstruction process of the decoder, the present invention provides an adaptive feature fusion mechanism to combine the multi-scale features provided by the encoder. The adaptive feature fusion mechanism here refers to introducing a jump connection with the corresponding layer of the encoder in each layer of the decoder, and directly introducing the deep features in the encoder into the decoder, thereby improving the accuracy of reconstruction and the guarantee of the original data.

[0050] Furthermore, the decoder uses multi-resolution deconvolution kernels to adapt to the different feature scales in the cable cover sensor data. The deconvolution kernel size of each layer is different, ensuring that the decoder can capture data details at different resolutions during reconstruction. This multi-resolution design makes the decoder more flexible in processing complex signals and can accurately reconstruct various features of the signal. The last layer of the decoder is a fully connected layer, which adjusts the final output of the decoder to the same dimension as the original input data of the sensor. Assume that the output dimension of the fully connected layer is , where N is the number of sensors and L is the number of sampling points of each sensor. The output dimension of the fully connected layer is matched with the number of sensors and the number of sampling points of each sensor, thereby ensuring that the reconstructed data corresponds seamlessly to the original data and effectively maintaining the structural integrity of the reconstructed data.

[0051] In this embodiment, in the reconstruction process of the latent variables, a multi-dimensional recovery mechanism is also introduced. The mean and variance in the latent space are combined together for reconstruction. The reconstruction formula can be expressed as: in, To reconstruct the data, is the mean of the latent variable, generated by the encoder, is the variance of the latent variable, which is used to capture the uncertainty of the data, is sampling noise, used to introduce randomness in the latent space, is the deconvolution function of the decoder.

[0052] In addition, in this embodiment, the decoder uses batch normalization and regularization mechanisms to prevent overfitting of data during the reconstruction process and ensure that the output data and the input sensor signal remain consistent. Through the normalization operation, the decoder can adapt to different operating conditions and sensor noise levels of the cable cover and generate high-fidelity reconstructed data. While improving the reconstruction quality, the consistency between the output data and the input sensor signal is ensured.

[0053] In a preferred embodiment, the loss function of the data reconstruction model adopts a weighted function based on the reconstruction error loss and the latent space regularization loss. Specifically, the weighted loss function Including reconstruction error loss and latent space regularization loss , by setting the dynamic weight coefficient To achieve balance, the loss function is expressed as: in, is the reconstruction error loss, which is used to measure the difference between the model output and the original sensor data. is the latent space regularization loss, which is used to ensure that the distribution in the latent space conforms to the preset prior distribution. It is a dynamic weight coefficient, which is used to summarize the adaptive adjustment according to the performance of the model during the training process to adapt to the feature distribution of the cable cover under different operating conditions.

[0054] Specifically, the regularization of the latent space uses an adaptive regularization method based on information entropy to independently evaluate each latent variable z in the latent space. The formula of its loss function is as follows: in, is a latent variable The information entropy of The distribution complexity of For the prior distribution and the true distribution The KL divergence is used to measure the difference between the two. is the weight coefficient of entropy regularization, which ensures that the distribution of the latent space is as simple as possible and avoids overly complex distribution structures.

[0055] The reconstruction error uses multi-scale feature matching loss to compare the reconstructed data with the original data at different scales to ensure that the global features of the data are taken into account while not neglecting the local details during the reconstruction process. The reconstruction error loss includes error metrics at different resolutions: in, and Respectively represent the identities of the original data and the reconstructed data at the sth scale, The weights for each scale ensure that the global features of the data are taken into account while local details are not neglected during the reconstruction process.

[0056] The dynamic weight coefficient can be adaptively adjusted according to the reconstruction error and the change of the potential space distribution, by calculating the reconstruction error change rate in each iteration and the rate of change of the potential spatial distribution , to dynamically adjust the weights: in, is the difference between the reconstruction error of the current batch and the previous batch, that is, the rate of change of the potential space distribution, is the difference between the regularization error of the current batch and the previous batch, that is, the rate of change of the reconstruction error, is the weight coefficient before adjustment, is the adjusted weight coefficient.

[0057] When the reconstruction error is large, the model automatically tends to increase The weight of , and when the latent space distribution deviates greatly from the expected prior, it tends to increase The weights of the model are adjusted to achieve adaptive tuning of model performance by balancing reconstruction error and latent space regularization loss, thereby maintaining efficient anomaly detection and data reconstruction capabilities in a changing environment. The model's focus on reconstruction accuracy and latent space regularization is flexibly adjusted according to the complexity of the data and the consistency of the latent space through the adjustment of dynamic weight coefficients, thereby improving the adaptability of the model.

[0058] In this embodiment, for the data reconstruction model of the above structure, a dynamic learning rate adjustment strategy is adopted during the training process, and the learning rate is dynamically adjusted to avoid the model from converging too slowly in the early stage of training or from oscillating in the later stage. The dynamic learning rate adjustment strategy in this embodiment includes a dynamic adjustment strategy based on the loss change rate and the gradient norm and a periodic dynamic adjustment strategy. The two adjustment strategies are described below.

[0059] When a dynamic adjustment strategy based on the loss change rate and gradient norm is adopted, the model calculates the rate of change of the loss function at each iteration and the norm of the gradient , as the basis for adjusting the learning rate: in, is the new learning rate, is the learning rate of the previous iteration, is the loss change rate between the two iterations, is the norm of the current gradient, A control parameter for learning rate adjustment, used to control the sensitivity of learning rate updates.

[0060] In this adjustment strategy, the new learning rate is determined by the learning rate of the previous iteration, the loss change rate and the norm of the current gradient. By controlling the sensitivity of the learning rate update, a more efficient learning process can be achieved.

[0061] When the periodic dynamic adjustment strategy is adopted, the learning rate is updated according to the current number of training steps, the total number of steps in the learning rate cycle, and the extreme value of the learning rate: in, represents the new learning rate, and Respectively represent the minimum and maximum values ​​of the learning rate, d represents the current number of training steps, Represents the total number of steps in the learning rate cycle.

[0062] The periodic dynamic adjustment strategy sets the minimum and maximum values ​​of the learning rate, and dynamically adjusts the learning rate according to the current number of training steps and the total number of steps in the learning rate cycle, thereby ensuring that the model always maintains an appropriate learning rate during the training process, enabling it to flexibly respond to changes in cable cover sensor data, improve the convergence speed of the model, prevent overfitting, and improve overall performance.

[0063] Furthermore, in another preferred embodiment, the present invention also adopts a data enhancement strategy to train the model, and the specific steps include: Acquire historical data of the cable cover plate collected by the sensor, and perform frequency domain disturbance enhancement and time domain random masking enhancement on the historical data to obtain first enhanced historical data and second enhanced historical data; A data set is constructed according to the historical data, the first enhanced historical data and the second enhanced historical data, and the data set is used to train the data reconstruction model.

[0064] In this embodiment, in order to enhance the adaptability of the model to the cable cover sensor data, an adaptive data enhancement strategy based on the data periodicity characteristics is adopted. Among them, the data enhancement strategy includes frequency domain perturbation enhancement and time domain random masking enhancement. Specifically, the frequency domain perturbation enhancement performs a discrete Fourier transform on the input signal, introduces a small perturbation with random amplitude and phase in the frequency domain, and then generates enhanced data back to the time domain through an inverse Fourier transform. Time domain random masking enhancement randomly selects several continuous segments of the time series data of the input signal to mask or fill noise, thereby simulating the data of the sensor in a bad state to be true or abnormal, and improving the model's ability to recognize abnormal data.

[0065] In this embodiment, new data samples are automatically generated by analyzing the periodic fluctuations of sensor data, thereby ensuring that the model has higher robustness in different environments.

[0066] After reconstructing the original data through the data reconstruction model to obtain reconstructed data, the cable cover abnormality detection can be performed based on the error change of the reconstructed data. The specific steps include: Obtaining a reconstruction error change rate according to a reconstruction error between the original data and the reconstructed data, comparing the reconstruction error change rate with an abnormal threshold, and obtaining abnormal data according to a comparison result; According to the duration and degree of abnormality of the abnormal data, an alarm reminder of a corresponding level is generated.

[0067] In this embodiment, real-time anomaly detection is performed by monitoring the reconstruction error. When the reconstruction error exceeds the set threshold, the signal will be marked as an abnormal signal and the corresponding alarm mechanism will be triggered to ensure real-time monitoring and response to the cable cover status. Specifically, the statistical distribution of the reconstruction error is dynamically calculated based on the raw data collected in real time, and the change trend of the reconstruction error is accumulated to obtain the reconstruction error change rate: in, is the cumulative rate of change of reconstruction error within the time window n, is the i-th reconstruction error. Within the time window n, the starting number is jn and the ending number is j. The changes of n reconstruction errors are counted and their average value is taken as the reconstruction error change rate. When a certain threshold is exceeded, an abnormal flag is triggered.

[0068] In this embodiment, by calculating the cumulative reconstruction error change rate, the progressive anomaly of the sensor signal can be captured, reducing false alarms caused by accidental error fluctuations. Preferably, the abnormal threshold can adopt a dynamic setting mechanism, such as using the mean and standard deviation in a local time window to set the dynamic threshold. This dynamic threshold adaptive mechanism enhances the robustness of the system, and can continuously and effectively monitor and evaluate the performance of the sensor under various working conditions, and respond to potential abnormal situations in a timely manner. Of course, the abnormal threshold can also be set manually or in other ways, which will not be repeated here.

[0069] In this embodiment, a multi-level alarm mechanism is adopted for the identified abnormal data. Specifically, the abnormal data is classified according to the duration and the degree of abnormality, wherein the degree of abnormality refers to the amplitude of the reconstruction error exceeding the dynamic threshold. Different levels of alarm responses are set for different types of abnormal data. For example, for abnormal data in which the reconstruction error exceeds the threshold in a short period of time but quickly returns to normal, a first-level alarm, i.e., a mild alarm, is triggered to alert the operation and maintenance personnel; for abnormal data in which the reconstruction error continues to exceed the threshold and the abnormal amplitude is large, a second-level alarm, i.e., a moderate alarm, is triggered, and the data is automatically saved for subsequent analysis; for abnormal data in which the reconstruction error exceeds the threshold for too long and the abnormal amplitude is extremely large, a third-level alarm, i.e., an emergency alarm, is triggered to promptly notify relevant personnel for emergency processing.

[0070] The present embodiment provides a cable cover abnormality detection method based on a variational autoencoder. The present invention can extract implicit features in multi-source data through the generative modeling capability of the variational autoencoder, automatically realize data reconstruction and fusion, enhance the status monitoring of the cable cover, thereby improving the integrity and reliability of the data, and at the same time utilize the unsupervised learning characteristics of the variational autoencoder and dynamically identify abnormal situations through reconstruction errors, which not only improves the flexibility and accuracy of detection, but also can respond to changes in the operating status of the cable cover in a timely manner, reduces the interference caused by human factors, significantly improves the real-time and effectiveness of abnormality identification, further reduces the workload of manual inspection and data analysis, improves operation and maintenance efficiency, and effectively extends the service life of the cable cover.

[0071] See also Figure 2 Based on the same inventive concept, a cable cover abnormality detection system based on a variational autoencoder is proposed in the second embodiment of the present invention, comprising: The data sampling module 10 is used to perform denoising on the original data of the cable cover plate collected by the sensor to obtain a denoised signal, and dynamically resample the denoised signal according to the local change rate to obtain resampled data; A data reconstruction module 20 is used to input the resampled data into a pre-established data reconstruction model to obtain reconstructed data, wherein the data reconstruction model is constructed based on a variational autoencoder and trained based on a dynamic learning rate adjustment strategy and a data enhancement strategy; The data reconstruction model consists of an encoder and a corresponding decoder. The encoder is constructed based on a deep convolutional neural network and uses a dynamic adjustment mechanism to adjust the mean and variance of the generated latent variables. The decoder consists of a deconvolution layer and uses an adaptive feature fusion mechanism and a multi-dimensional recovery mechanism to reconstruct the latent variables. The anomaly detection module 30 is used to perform anomaly detection according to the reconstruction error change rate of the reconstructed data to obtain abnormal data, and generate corresponding alarm reminders according to the abnormal data.

[0072] The technical features and technical effects of the cable cover anomaly detection system based on variational autoencoder proposed in the embodiment of the present invention are the same as the method proposed in the embodiment of the present invention, and will not be repeated here. Each module in the above-mentioned cable cover anomaly detection system based on variational autoencoder can be implemented in whole or in part by software, hardware and a combination thereof. The above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory in the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.

[0073] In summary, an embodiment of the present invention proposes a cable cover anomaly detection method and system based on a variational autoencoder. The method denoises the original data of the cable cover collected by the sensor to obtain a denoised signal, and dynamically resamples the denoised signal according to the local change rate to obtain resampled data; the resampled data is input into a pre-established data reconstruction model to obtain reconstructed data, and the data reconstruction model is constructed based on a variational autoencoder and trained based on a dynamic learning rate adjustment strategy and a data enhancement strategy; anomaly detection is performed according to the reconstruction error change rate of the reconstructed data to obtain abnormal data, and corresponding alarm reminders are generated according to the abnormal data; wherein the data reconstruction model is composed of an encoder and a corresponding decoder, the encoder is constructed based on a deep convolutional neural network, and a dynamic adjustment mechanism is used to adjust the mean and variance of the generated latent variables, the decoder is composed of a deconvolution layer, and an adaptive feature fusion mechanism and a multi-dimensional recovery mechanism are used to reconstruct the latent variables. Through the generative modeling capability of the variational autoencoder, the present invention can effectively extract implicit features in multi-source data, automatically realize data reconstruction and fusion, enhance the status monitoring of cable covers, thereby improving the integrity and reliability of the data, and at the same time utilize the unsupervised learning characteristics of the variational autoencoder and dynamically identify abnormal situations through reconstruction errors, which not only improves the flexibility and accuracy of detection, but also can promptly respond to changes in the operating status of the cable cover, reduces the interference caused by human factors, significantly improves the real-time and effectiveness of abnormality identification, further reduces the workload of manual inspections and data analysis, improves operation and maintenance efficiency, and effectively extends the service life of the cable cover.

[0074] Each embodiment in this specification is described in a progressive manner, and the same or similar parts of each embodiment can be directly referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the system embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment. It should be noted that the technical features of the above-mentioned embodiments can be combined arbitrarily. In order to make the description concise, all possible combinations of the technical features in the above-mentioned embodiments are not described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0075] The above-mentioned embodiments only express several preferred implementation modes of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for ordinary technicians in the technical field, several improvements and substitutions can be made without departing from the technical principles of the present invention, and these improvements and substitutions should also be regarded as the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be based on the protection scope of the claims.

Claims

1. A cable cover abnormality detection method based on variational autoencoder, characterized in that: The method comprises: De-noising the original data of the cable cover plate collected by the sensor to obtain a de-noised signal, and dynamically re-sampling the de-noised signal according to the local change rate to obtain re-sampled data; Inputting the resampled data into a pre-established data reconstruction model to obtain reconstructed data, wherein the data reconstruction model is constructed based on a variational autoencoder and trained based on a dynamic learning rate adjustment strategy and a data enhancement strategy; Performing anomaly detection according to a reconstruction error change rate of the reconstructed data to obtain abnormal data, and generating a corresponding alarm reminder according to the abnormal data; Among them, the data reconstruction model consists of an encoder and a corresponding decoder. The encoder is constructed based on a deep convolutional neural network, and a dynamic adjustment mechanism is used to adjust the mean and variance of the generated latent variables. The decoder is composed of a deconvolution layer, and an adaptive feature fusion mechanism and a multidimensional recovery mechanism are used to reconstruct the latent variables.

2. The cable cover plate abnormality detection method based on variational autoencoder according to claim 1 is characterized in that: The step of performing denoising processing on the raw data of the cable cover plate collected by the sensor to obtain a denoised signal comprises: Collecting raw data of the cable cover through sensors, wherein the raw data includes temperature data, humidity data and pressure data; Performing time-frequency conversion on the original data by Fourier transform to obtain a frequency domain signal, calculating the energy spectrum density of the frequency domain signal, comparing the energy spectrum density with a density threshold, and filtering interference signals on the frequency domain signal according to the comparison result to obtain a first frequency domain signal, wherein the density threshold is determined by the median and standard deviation of the energy spectrum density; The first frequency domain signal is frequency filtered using a frequency filter constructed based on a window function to obtain a second frequency domain signal, and the second frequency domain signal is time-frequency inverted to obtain a denoised signal.

3. The cable cover plate abnormality detection method based on variational autoencoder according to claim 1 is characterized in that: The step of dynamically resampling the denoised signal according to the local change rate to obtain resampled data comprises: Dividing the denoised signal into a plurality of time series intervals, and calculating the local change rate of the denoised signal according to the signal difference between adjacent sampling points in each time series interval; Comparing the local change rate with a change rate threshold, and determining a data sampling mode for each sensor according to the comparison result, wherein the data sampling mode includes a high-frequency sampling mode and a low-frequency sampling mode; Determine a resampling weight of each sensor based on historical data of each sensor, and dynamically adjust the sampling frequency in the data sampling mode according to the resampling weight to obtain a dynamic resampling frequency; According to the dynamic resampling frequency, the denoised signal corresponding to each sensor is dynamically resampled to obtain resampled data.

4. The cable cover plate abnormality detection method based on variational autoencoder according to claim 1 is characterized in that: The encoder includes multiple convolutional layers, multiple pooling layers and a first fully connected layer, the number of convolutional layers and pooling layers is the same, the convolution kernel sizes used in each convolutional layer are different, the output data of each convolutional layer is the input data of the corresponding pooling layer, the output data of the corresponding pooling layer is the input data of the next convolutional layer, and the output data of the last pooling layer is the input data of the first fully connected layer; The encoder uses a dynamic adjustment mechanism based on the rate of change to adjust the mean and variance of the generated latent variables; The decoder includes a plurality of sequentially connected deconvolution layers and a second fully connected layer, the deconvolution kernel sizes used in the various deconvolution layers are different, the last deconvolution layer is connected to the second fully connected layer, the number of deconvolution layers is consistent with the number of convolution layers of the encoder, and each deconvolution layer and the corresponding convolution layer use a shared weight matrix; The encoder and the decoder have a skip connection structure, and the decoder reconstructs the latent variables using a multi-dimensional recovery mechanism combining the mean and the variance.

5. The cable cover plate abnormality detection method based on variational autoencoder according to claim 4 is characterized in that: The loss function of the data reconstruction model is a weighted function based on the reconstruction error loss and the potential space regularization loss, and the weight coefficient in the loss function is a dynamic weight coefficient that is dynamically adjusted based on the reconstruction error change rate and the potential space distribution change rate; The loss function is expressed by the following formula: In the formula, represents the reconstruction error loss, represents the latent space regularization loss, represents the weight coefficient; The weight coefficient is expressed by the following formula: In the formula, represents the adjusted weight coefficient, represents the weight coefficient before adjustment, represents the rate of change of reconstruction error, Represents the rate of change of the potential spatial distribution.

6. The cable cover plate abnormality detection method based on variational autoencoder according to claim 5 is characterized in that: The steps of training the data reconstruction model based on a dynamic learning rate adjustment strategy include: Calculating the rate of change and the gradient norm of the loss function, updating the learning rate according to the rate of change and the gradient norm, and iteratively training the data reconstruction model using the updated learning rate; The learning rate is expressed by the following formula: In the formula, represents the new learning rate, represents the learning rate of the previous iteration, represents the loss change rate, represents the gradient norm, Indicates control parameters.

7. The cable cover plate abnormality detection method based on variational autoencoder according to claim 5 is characterized in that: The steps of training the data reconstruction model based on a dynamic learning rate adjustment strategy include: The learning rate is updated according to the current number of training steps, the total number of steps in the learning rate cycle and the extreme value of the learning rate, and the updated learning rate is used to iteratively train the data reconstruction model; The learning rate is expressed by the following formula: in, represents the new learning rate, and Respectively represent the minimum and maximum values ​​of the learning rate, d represents the current number of training steps, Represents the total number of steps in the learning rate cycle.

8. The cable cover plate abnormality detection method based on variational autoencoder according to claim 1 is characterized in that: The steps of training the data reconstruction model based on the data enhancement strategy include: Acquire historical data of the cable cover plate collected by the sensor, and perform frequency domain disturbance enhancement and time domain random masking enhancement on the historical data to obtain first enhanced historical data and second enhanced historical data; A data set is constructed according to the historical data, the first enhanced historical data and the second enhanced historical data, and the data set is used to train the data reconstruction model.

9. The cable cover plate abnormality detection method based on variational autoencoder according to claim 1 is characterized in that: The step of performing anomaly detection according to the reconstruction error change rate of the reconstructed data to obtain abnormal data, and generating a corresponding alarm reminder according to the abnormal data includes: Obtaining a reconstruction error change rate according to a reconstruction error between the original data and the reconstructed data, comparing the reconstruction error change rate with an abnormal threshold, and obtaining abnormal data according to a comparison result; According to the duration and degree of abnormality of the abnormal data, an alarm reminder of a corresponding level is generated.

10. A cable cover abnormality detection system based on variational autoencoder, characterized in that: The system comprises: The data sampling module is used to perform denoising on the original data of the cable cover plate collected by the sensor to obtain a denoised signal, and dynamically resample the denoised signal according to the local change rate to obtain resampled data; A data reconstruction module, used for inputting the resampled data into a pre-established data reconstruction model to obtain reconstructed data, wherein the data reconstruction model is constructed based on a variational autoencoder and trained based on a dynamic learning rate adjustment strategy and a data enhancement strategy; The data reconstruction model consists of an encoder and a corresponding decoder. The encoder is constructed based on a deep convolutional neural network and uses a dynamic adjustment mechanism to adjust the mean and variance of the generated latent variables. The decoder consists of a deconvolution layer and uses an adaptive feature fusion mechanism and a multi-dimensional recovery mechanism to reconstruct the latent variables. The anomaly detection module is used to perform anomaly detection according to the reconstruction error change rate of the reconstructed data to obtain abnormal data, and generate corresponding alarm reminders according to the abnormal data.

Images